index: codedb.snapshot written into project root pollutes git / risks committing a multi-MB index

[justrach]

Summary

The codedb MCP server exposes a large tool surface (~23 tools). With Claude Code's deferred-tool mechanism this means every session spends ToolSearch round-trips just to load tool schemas before doing any real work, and the tool definitions tax cached context each turn.

Evidence

• Every audited agent run began with 2 ToolSearch calls (to load codedb tool schemas) before the first real action.
• In the A/B benchmark, codedb runs emitted +23.5% more output tokens than the native-tools baseline (more, smaller tool calls), which is the main reason a ~23% raw-token saving converted to only ~2% cost saving at Sonnet pricing.

Impact

• Per-session schema-loading overhead and per-turn context tax.
• High tool-call/output-token count blunts the cost benefit of the token savings.

Suggested direction

• Consolidate overlapping tools (e.g. read/outline; search/find/word/glob) into fewer, mode-flagged tools.
• Ship a slim default toolset, with the long tail opt-in.

Found via an independent SWE-bench Lite token-efficiency benchmark: identical agent (`claude -p`, Sonnet 4.6) and tasks, only the tool surface differs - native Read/Grep/Edit vs codedb MCP tools. N=51 paired instances; full harness + data available.

[justrach]

Fixed in c7bce2b: after writing the in-tree snapshot, codedb.snapshot is appended to .git/info/exclude (local, untracked — leaves your .gitignore alone). Best-effort + idempotent; only the in-tree write triggers it. Merged to release/0.2.5825 and release/0.2.5826; PR #631 → main. Test: issue-625: in-tree snapshot is added to .git/info/exclude.